Projects · Open data

Open-model lineage

Which open models does everyone build on? Each point is a model on Hugging Face, joined to the model it was fine-tuned, merged or adapted from, and to the lab that released it.

Method

Start from the 500 most-downloaded text-generation and image-text-to-text models of the last 30 days.
Follow the base_model each repository declares, up to a model with no parent. Undeclared parents are not guessed.
Add the most-downloaded children of the largest families and of every base model above 100K downloads.
Quantised builds (GGUF, AWQ, FP8 and the like) carry no lineage of their own, so they are counted on the model they compress instead of drawn.
Leave out CI test fixtures and repositories with fewer than five likes.
Point size is downloads over the last 30 days, so the map shows what people use now, not what mattered first.

Source: the public Hugging Face Hub API.